IVGCVSW-2212 Add Neon support for Maximum operator

Change-Id: Iddae3486641d2d195827ae4e0a9dfa3f7ccd65e3

9 files changed, 106 insertions, 1 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index 869fd03251..724455de40 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -23,6 +23,7 @@
 #include "workloads/NeonConvolution2dWorkload.hpp"
 #include "workloads/NeonDepthwiseConvolutionWorkload.hpp"
 #include "workloads/NeonL2NormalizationFloatWorkload.hpp"
+#include "workloads/NeonMaximumWorkload.hpp"
 #include "workloads/NeonMergerWorkload.hpp"
 #include "workloads/NeonMultiplicationFloatWorkload.hpp"
 #include "workloads/NeonNormalizationFloatWorkload.hpp"
@@ -346,6 +347,18 @@ bool NeonLayerSupport::IsLstmSupported(const TensorInfo& input,
     return false;
 }
 
+bool NeonLayerSupport::IsMaximumSupported(const TensorInfo& input0,
+                                          const TensorInfo& input1,
+                                          const TensorInfo& output,
+                                          Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonMaximumWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input0,
+                                   input1,
+                                   output);
+}
+
 bool NeonLayerSupport::IsMeanSupported(const TensorInfo& input,
                                        const TensorInfo& output,
                                        const MeanDescriptor& descriptor,
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index 43d0bd9a81..2871c795f0 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -121,6 +121,11 @@ public:
                          const TensorInfo* cellToOutputWeights,
                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsMaximumSupported(const TensorInfo& input0,
+                            const TensorInfo& input1,
+                            const TensorInfo& output,
+                            Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsMeanSupported(const TensorInfo& input,
                          const TensorInfo& output,
                          const MeanDescriptor& descriptor,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index f7b2133196..848a6f397b 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -267,7 +267,7 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp32ToFp16(
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMaximum(const MaximumQueueDescriptor& descriptor,
                                                               const WorkloadInfo& info) const
 {
-    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
+    return std::make_unique<NeonMaximumWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor,
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index 88fb643c64..a3058ad73e 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -25,6 +25,7 @@ BACKEND_SOURCES := \
         workloads/NeonFullyConnectedWorkload.cpp \
         workloads/NeonL2NormalizationFloatWorkload.cpp \
         workloads/NeonLstmFloatWorkload.cpp \
+        workloads/NeonMaximumWorkload.cpp \
         workloads/NeonMergerWorkload.cpp \
         workloads/NeonMultiplicationFloatWorkload.cpp \
         workloads/NeonNormalizationFloatWorkload.cpp \
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 72ae3728d0..37933e0254 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -399,6 +399,14 @@ ARMNN_AUTO_TEST_CASE(LstmLayerFloat32NoCifgNoPeepholeNoProjection,
 ARMNN_AUTO_TEST_CASE(LstmLayerFloat32NoCifgWithPeepholeWithProjection,
                      LstmLayerFloat32NoCifgWithPeepholeWithProjectionTest)
 
+// Max
+ARMNN_AUTO_TEST_CASE(SimpleMaximum, MaximumSimpleTest)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1Element, MaximumBroadcast1ElementTest)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1DVector, MaximumBroadcast1DVectorTest)
+ARMNN_AUTO_TEST_CASE(MaximumUint8, MaximumUint8Test)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1ElementUint8, MaximumBroadcast1ElementUint8Test)
+ARMNN_AUTO_TEST_CASE(MaximumBroadcast1DVectorUint8, MaximumBroadcast1DVectorUint8Test)
+
 // Normalization
 ARMNN_AUTO_TEST_CASE(SimpleNormalizationAcross, SimpleNormalizationAcrossTest)
 ARMNN_AUTO_TEST_CASE(SimpleNormalizationWithin, SimpleNormalizationWithinTest)
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 2c2c9f05e6..9961397445 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -28,6 +28,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonL2NormalizationFloatWorkload.hpp
     NeonLstmFloatWorkload.cpp
     NeonLstmFloatWorkload.hpp
+    NeonMaximumWorkload.cpp
+    NeonMaximumWorkload.hpp
     NeonMergerWorkload.cpp
     NeonMergerWorkload.hpp
     NeonMultiplicationFloatWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonMaximumWorkload.cpp b/src/backends/neon/workloads/NeonMaximumWorkload.cpp
new file mode 100644
index 0000000000..9994af98d1
--- /dev/null
+++ b/src/backends/neon/workloads/NeonMaximumWorkload.cpp
@@ -0,0 +1,45 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonMaximumWorkload.hpp"
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+
+namespace armnn
+{
+
+arm_compute::Status NeonMaximumWorkloadValidate(const TensorInfo& input0,
+                                                const TensorInfo& input1,
+                                                const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInput0 = armcomputetensorutils::BuildArmComputeTensorInfo(input0);
+    const arm_compute::TensorInfo aclInput1 = armcomputetensorutils::BuildArmComputeTensorInfo(input1);
+    const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    return arm_compute::NEElementwiseMax::validate(&aclInput0,
+                                                   &aclInput1,
+                                                   &aclOutput);
+}
+
+NeonMaximumWorkload::NeonMaximumWorkload(const MaximumQueueDescriptor& descriptor,
+                                         const WorkloadInfo& info)
+    : BaseWorkload<MaximumQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonMaximumWorkload", 2, 1);
+
+    arm_compute::ITensor& input0 = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& input1 = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+    arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    m_MaxLayer.configure(&input0, &input1, &output);
+}
+
+void NeonMaximumWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonMaximumWorkload_Execute");
+    m_MaxLayer.run();
+}
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonMaximumWorkload.hpp b/src/backends/neon/workloads/NeonMaximumWorkload.hpp
new file mode 100644
index 0000000000..59e2d6104d
--- /dev/null
+++ b/src/backends/neon/workloads/NeonMaximumWorkload.hpp
@@ -0,0 +1,30 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <neon/workloads/NeonWorkloadUtils.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEElementwiseOperations.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonMaximumWorkloadValidate(const TensorInfo& input0,
+                                                const TensorInfo& input1,
+                                                const TensorInfo& output);
+
+class NeonMaximumWorkload : public BaseWorkload<MaximumQueueDescriptor>
+{
+public:
+    NeonMaximumWorkload(const MaximumQueueDescriptor& descriptor, const WorkloadInfo& info);
+
+    virtual void Execute() const override;
+
+private:
+    mutable arm_compute::NEElementwiseMax m_MaxLayer;
+};
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 48bd137560..82142f2aa1 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -16,6 +16,7 @@
 #include "NeonFullyConnectedWorkload.hpp"
 #include "NeonL2NormalizationFloatWorkload.hpp"
 #include "NeonLstmFloatWorkload.hpp"
+#include "NeonMaximumWorkload.hpp"
 #include "NeonMergerWorkload.hpp"
 #include "NeonMultiplicationFloatWorkload.hpp"
 #include "NeonNormalizationFloatWorkload.hpp"